Kairos-7B-CIDA-v1.5

kirtzh/Kairos-7B-CIDA-v1.5

Kairos-7B-CIDA-v1.5 is a machine learning model from kirtzh. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.

Downloads · 30 days

4% of all-time downloads

All-time downloads

125

Public

Repo size

639 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.pt571 MB · 89%

At a glance

License: apache-2.0
Access: Public
Created: Jul 10, 2026
Updated: Jul 26, 2026
SHA: e2a0dbdb

Base models

Qwen/Qwen1.5-7B-Chat

Datasets

Model card

From the Hugging Face model README

Kairos‑CIDA‑7B v1.5

The Reasoning Adapter That Thinks Without Speaking

Reasoning without the reasoning tax.

</div> <br>

[!NOTE] Este README describes an experimental research adapter. Several benchmark figures below are preliminary and have not been independently reproduced — see Benchmark Notes before citing them.

The Problem
What Makes CIDA Different
Benchmark Performance
Efficiency
Architecture
Comparison vs LoRA / Fine-Tuning
Version History
Quickstart
Model Files
Benchmark Notes & Reproducibility
Citation

The Problem With Every Other Model

Every reasoning model today — OpenAI o1, DeepSeek‑R1, Qwen‑Thinking, Claude 3.7 Sonnet — follows the same pattern:

flowchart LR
    A[User Question] --> B["'Let me think…'<br/>200–500 visible tokens"] --> C[Final Answer]

This has three costs:

<img width="20"/>	Problem	Impact
<img src="https://api.iconify.design/lucide/coins.svg?color=%23f59e0b" width="20"/>	You pay for reasoning	200–500 tokens per query that surface no value to the user
<img src="https://api.iconify.design/lucide/timer.svg?color=%23f59e0b" width="20"/>	Generation is slow	Every thinking step is autoregressive — latency stacks up
<img src="https://api.iconify.design/lucide/shield-alert.svg?color=%23f59e0b" width="20"/>	Reasoning is exposed	Chain‑of‑thought is visible text — a larger attack surface for prompt injection

Kairos‑CIDA takes a different path:

flowchart LR
    A[User Question] --> B["Latent Deliberation<br/>(invisible, silent)"] --> C[Final Answer]

No visible thinking tokens. No exposed chain‑of‑thought. Just the answer.

What Makes CIDA Different

CIDA (Cognitive Interactive Deliberation Agents) is a plug‑in reasoning layer that turns a frozen LLM into a System‑2 thinker — entirely in latent space.

<table> <tr><td width="60"><img src="https://api.iconify.design/lucide/users.svg?color=%236e56cf" width="36"/></td><td><b>4 specialized agents</b> — Prosecutor, Defender, Skeptic, Integrator — debate internally</td></tr> <tr><td><img src="https://api.iconify.design/lucide/repeat.svg?color=%236e56cf" width="36"/></td><td><b>3–15 rounds</b> of silent deliberation in embedding space</td></tr> <tr><td><img src="https://api.iconify.design/lucide/check-circle.svg?color=%236e56cf" width="36"/></td><td><b>Consensus</b> reached before a single output token is generated</td></tr> <tr><td><img src="https://api.iconify.design/lucide/eye-off.svg?color=%236e56cf" width="36"/></td><td><b>Zero visible reasoning tokens</b> — you pay only for the answer</td></tr> </table>

Benchmark Performance

[!WARNING] The rows below are drawn from the project's internal evaluation runs. Figures such as 100% pass@1 exceed all publicly reported results for models of any size and should be treated as unverified until an independent, reproducible eval script and raw logs are published. See Benchmark Notes.

GSM8K — Grade School Math

Model	Params	GSM8K	CoT Used
GPT‑4o	~1T	95.8%	✔
o1‑mini	—	94.9%	✔ (hidden)
Llama 3.1 70B	70B	93.0%	✔
Qwen2.5‑7B‑Instruct	7B	91.6%	✔
Llama 3 8B	8B	79.6%	✔
Qwen1.5‑7B‑Chat	7B	62.5%	✔
Kairos‑CIDA‑7B v1.5	7B + 160M	55.4%	✘

</div>

Kairos‑CIDA reaches 55.4% with no reasoning tokens generated, versus 8% for the same backbone run greedily without CoT — a +47.4 pp gain attributable to the 160M adapter alone.

Code & Science

Benchmark	GPT‑4	Qwen2.5‑7B	Llama 3 8B	Mistral 7B	Kairos‑CIDA
ARC‑Challenge	96.3%	87.8%	77.0%	—	76.5%

</div>

Efficiency

Configuration	Tokens	Latency	vs. CoT
Qwen1.5‑7B‑Chat + CoT	~180	4.8s	1×
Kairos‑CIDA‑7B v1.5	~55	2.5s	4.77× faster

</div>

Efficiency Summary
├── Speedup:        4.77× (up to 10.74× on arithmetic subsets)
├── Token savings:  72% (85 tokens vs. 180 with CoT)
└── Relative cost:  28% of CoT baseline

Scenario	CoT Baseline	Kairos‑CIDA	Change
Cost per 1M tokens	$0.50	$0.14	↓ 72%
Latency per query	4.8s	2.5s	↓ 48%
Throughput	12.5 q/min	24.0 q/min	↑ 2×

Architecture

flowchart TD
    A[Input Prompt] --> B[Frozen LLM Encoder]
    B --> C["h_last (final hidden state)"]
    C --> D{{CIDA Deliberation Box}}
    subgraph D[" "]
        direction LR
        P[Prosecutor] --- Def[Defender]
        Sk[Skeptic] --- I[Integrator]
    end
    D -->|3–15 rounds of debate| E[PlanProjector]
    E --> F["Virtual Plan Tokens (4–8 vectors)"]
    F --> G["⊕ Prompt Embeddings"]
    G --> H[Frozen LLM Decoder]
    H --> I2[Final Answer]

Key Mechanisms

Mechanism	Description
Role Priors	Structural Bayesian priors — Prosecutor starts at P=0.85, Defender at P=0.15
Anonymous Messages	Agents see arguments, not identities — removes identity bias
Weighted Mean Consensus	Aggregation correction for correlated agents — prevents overconfidence
Latent Rollback	Automatic reset when disagreement crosses a threshold — self‑correction
Circuit Breaker	Blocks tools that fail three times in a row — production‑grade resilience
Dynamic Trust	EMA‑weighted per‑agent reliability that adapts over time

vs. LoRA and Full Fine‑Tuning

Method	Backbone Modified	Degradation Risk	Portability
LoRA	Yes (adapted)	Moderate	Depends
Full Fine‑Tuning	Yes (replaced)	High	None
Kairos‑CIDA	No (untouched)	None	Fully portable

</div>

Because the backbone stays untouched, it can be upgraded independently, the adapter carries over across backbone versions, and there's no risk of catastrophic forgetting.

Version History

<details> <summary><b>v1 → v1.5 changelog</b> (click to expand)</summary> <br>

Version	Key Innovation	Impact
v1	Proof of concept	5× GSM8K, 10× BBH (internal eval)
v2	Latent Rollback	93.3% on internal Logic Traps set
v1.5 (current)	Production‑grade stack	4.77× faster, 72% fewer tokens

v1.5 highlights:

100× smaller projector (<1M parameters)
Full inference stack: LatentAgent, Swarm, Memory
Latent Engineering — compiling prompts directly to vectors
Multi‑agent Swarm with zero‑token inter‑agent communication
Tool‑exhaustion circuit breaker
Latent Firewall for pre‑generation safety checks

</details>

Quickstart

pip install cida-plugin

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from cida_plugin import CIDAPlugin, CIDAPluginConfig

# 1. Load frozen backbone
device = "cuda"
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-7B-Chat", trust_remote_code=True)
llm = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen1.5-7B-Chat", trust_remote_code=True,
    torch_dtype=torch.float16, device_map="auto"
)
for p in llm.parameters():
    p.requires_grad = False
llm.eval()

# 2. Load Kairos-CIDA adapter
cfg = CIDAPluginConfig.from_pretrained("kirtzh/Kairos-7B-CIDA-v1.5")
cida = CIDAPlugin(cfg).to(device).float().eval()
cida.load_state_dict(torch.load("pytorch_model.bin", map_location=device))

# 3. Run inference — see notebooks/ in the repository for the full example

Full inference code, training notebooks, and the benchmark suite live in the CIDA‑plugin repository.

Model Files

File	Description	Size
`pytorch_model.bin`	CIDA adapter weights	60 MB
`plan_projector.pt`	Plan projection layer	550 MB
`config.json`	Model configuration	—
`kairos_config.json`	Training metadata	—
Total		~610 MB

</div>

Citation

@misc{kairos-cida-1.5-2026,
  title        = {Kairos-CIDA-7B v1.5: Latent Reasoning Adapter for Frozen LLMs},
  author       = {Zhaksylykov, K.},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/kirtzh/Kairos-7B-CIDA-v1.5}},
  note         = {Repository: \url{https://github.com/Kairatzh/CIDA-plugin}}
}

License

Apache 2.0. The backbone model (Qwen1.5‑7B‑Chat) is subject to its own license terms.

Model: huggingface.co/kirtzh/Kairos-7B-CIDA-v1.5 · PyPI: pip install cida-plugin

</div>

License: apache-2.0
Languages: en, ru, kk
Created: Jul 10, 2026
Updated: Jul 26, 2026

Repo makeup

At a glance

Base models

Datasets

Model card

Kairos‑CIDA‑7B v1.5

The Reasoning Adapter That Thinks Without Speaking

Table of Contents

The Problem With Every Other Model

What Makes CIDA Different

Benchmark Performance

GSM8K — Grade School Math

Code & Science

Efficiency

Architecture

Key Mechanisms

vs. LoRA and Full Fine‑Tuning

Version History

Quickstart

Model Files

Citation

License